Unpaired Optical Coherence Tomography Angiography Image Super-Resolution via Frequency-Aware Inverse-Consistency GAN

TL;DR

This work tackles unpaired OCTA image super-resolution by introducing Frequency-aware Unpaired Super-Resolution (FUSR), which uses a dual-path generator to separately refine high-frequency ($hf$) capillary details and preserve low-frequency ($lf$) structure. It couples a frequency-aware adversarial loss (FAL) based on Haar wavelets with a frequency-aware focal consistency loss (FFCL) to enforce spectral fidelity during end-to-end training of restoration and degradation mappings, $G_{Res}$ and $G_{Deg}$. Evaluations on the CUHK-STDR dataset demonstrate superior quantitative metrics (e.g., PSNR, SSIM, FSIM, NMI) and clearer fine vasculature compared to state-of-the-art unpaired SR methods, with significant statistical improvements. The approach advances practical OCTA SR by reducing reliance on paired data while preserving diagnostically important microvascular details, enabling better assessment of retinal diseases in clinical settings. Areas for future work include self-supervised extensions and multi-device validation to improve generalizability.

Abstract

For optical coherence tomography angiography (OCTA) images, a limited scanning rate leads to a trade-off between field-of-view (FOV) and imaging resolution. Although larger FOV images may reveal more parafoveal vascular lesions, their application is greatly hampered due to lower resolution. To increase the resolution, previous works only achieved satisfactory performance by using paired data for training, but real-world applications are limited by the challenge of collecting large-scale paired images. Thus, an unpaired approach is highly demanded. Generative Adversarial Network (GAN) has been commonly used in the unpaired setting, but it may struggle to accurately preserve fine-grained capillary details, which are critical biomarkers for OCTA. In this paper, our approach aspires to preserve these details by leveraging the frequency information, which represents details as high-frequencies ($\textbf{hf}$) and coarse-grained backgrounds as low-frequencies ($\textbf{lf}$). In general, we propose a GAN-based unpaired super-resolution method for OCTA images and exceptionally emphasize $\textbf{hf}$ fine capillaries through a dual-path generator. To facilitate a precise spectrum of the reconstructed image, we also propose a frequency-aware adversarial loss for the discriminator and introduce a frequency-aware focal consistency loss for end-to-end optimization. Experiments show that our method outperforms other state-of-the-art unpaired methods both quantitatively and visually.

Figures (10)

• Figure 1: Illustration for our OCTA images dataset, which is retrospectively collected from the Chinese University of Hong Kong Sight-THReatening Diabetic Retinopathy (CUHK-STDR) study. Leftmost illustrates the original 6mm$\times$6mm image. Orange boxes with subscript 1 indicate 6mm$\times$6mm patches. Blue boxes with subscript 2 indicate the corresponding paired 3mm$\times$3mm images. The same letters indicate the same area. A: Fovea-center images. B$\sim$E: Parafoveal images.
• Figure 2: Illustration for 6mm$\times$6mm OCTA images and frequency components. Lower-right corners are bandwidth spectral filters. A: 6mm$\times$6mm OCTA image and its spectrum. B: $hf$ and its high-pass filter. C: $lf$ and its low-pass filter. D: Middle-frequencies and its middle-pass filter.
• Figure 3: Azimuthal integral on spectrum as specified in Eq. \ref{['azimuthal']}. It indicates that $HR$ contains stronger power in middle- and high-bands of the spectrum than $LR$. While the middle frequencies of different methods are similar, our approach better fits the $lf$ information and enhances $hf$ information compared to the real $HR$.
• Figure 4: An overview of our methods. The input $LR$ image is decomposed into $lf$ and $hf^*$ (through $HFB$) and fused for restoring to $HR^{\uparrow}$. Then, it is degraded to $LR^{\uparrow\downarrow}$ and optimized through restoration-degradation consistency and adversarial losses. The restoration network is eventually taken as the OCTA high-resolution model in the inference phase. The inverse degradation-restoration process is represented in simplified conceptual graphs. Note that input $LR$ and $HR$ images are unpaired in the training phase.
• Figure 5: Structure of the discriminator. To distinguish the image as either real or generated, our method combines both frequency and spatial information in the discriminating phase. Results are aggregated to formulate the frequency-aware adversarial loss.